Historical information-based Bayesian algorithm optimization method, apparatus and device, and medium

Abstract

The invention discloses a Bayesian algorithm optimization method and device based on historical information, equipment and a medium, and relates to the technical field of algorithm optimization, and the method comprises the steps: building an initial Gaussian process proxy model through obtaining an initial point set, generating a candidate point set through LHS, obtaining a collection function, and obtaining a hyper-parameter and a time decay factor corresponding to the collection function, and obtaining a Bayesian process proxy model; and obtaining a target function evaluation point in the candidate point set by combining the initial Gaussian process agent model with the acquisition function and the hyper-parameter corresponding to the acquisition function, and improving the initial Gaussian process agent model based on the target function evaluation point and the time decay factor to obtain a target optimization algorithm. The optimization efficiency and precision of the Bayesian algorithm are improved by continuously updating a GP model, combining a dynamic acquisition function and a hyper-parameter and a time decay factor of the dynamic acquisition function and iteratively optimizing an objective function evaluation point.

Description

Technical Field

[0001] The present invention relates to the technical field of algorithm optimization, and in particular to a Bayesian algorithm optimization method, device, equipment and medium based on historical information. Background Art

[0002] In the field of intelligent computing and machine learning, Bayesian Optimization (BO) is widely used in many fields, including but not limited to machine learning hyperparameter tuning, experimental design, industrial optimization, and dynamic system optimization. The workflow of BO usually involves building a proxy model (such as a Gaussian process) to model the objective function. For example, in experimental design, directly evaluating the objective function may require a time-consuming experimental process; while in machine learning hyperparameter tuning, it may require a lot of computing resources. These limitations make it difficult for traditional optimization methods to cope with them efficiently, and Bayesian optimization can effectively reduce the number of calls to the objective function through intelligent sampling strategies, thereby approaching the optimal solution under limited sampling conditions.

[0003] Traditional Bayesian optimization methods rely on static acquisition functions and fixed hyperparameter settings for optimization. However, due to the fixed acquisition functions and their hyperparameters, traditional Bayesian optimization methods are difficult to adapt to the needs of different stages in the optimization process, and cannot fully utilize the historical information in previous iterations. Therefore, there is an urgent need for a Bayesian optimization method based on historical information to improve the efficiency of optimization. Summary of the invention

[0004] The main purpose of this application is to provide a Bayesian algorithm optimization method, device, equipment and medium based on historical information, aiming to solve the technical problem of how to improve the efficiency of Bayesian algorithm optimization.

[0005] To achieve the above objectives, the present application proposes a Bayesian algorithm optimization method based on historical information, the method comprising:

[0006] Get the initial point set;

[0007] Constructing an initial Gaussian process proxy model according to the initial point set;

[0008] Generate a candidate point set through LHS, obtain an acquisition function, a hyperparameter corresponding to the acquisition function, and a time decay factor;

[0009] Obtaining an objective function evaluation point in a candidate point set by combining the initial Gaussian process proxy model with the acquisition function and a hyperparameter corresponding to the acquisition function;

[0010] The initial Gaussian process proxy model is improved based on the objective function evaluation point and the time decay factor to obtain a target optimization algorithm.

[0011] In one embodiment, the step of improving the initial Gaussian process proxy model based on the objective function evaluation point and the time decay factor to obtain a target optimization algorithm includes:

[0012] According to the objective function evaluation point, a true value of the objective function is obtained;

[0013] Based on the objective function evaluation point and the true value of the objective function, the training data set is expanded; based on the training data set and the time decay factor, the initial Gaussian process proxy model is improved to obtain a target optimization algorithm.

[0014] In one embodiment, the step of expanding the training data set based on the objective function evaluation point and the objective function true value includes:

[0015] According to the objective function evaluation point, the objective function is used to perform calculation to obtain the true value of the objective function corresponding to the objective function evaluation point;

[0016] Combining the objective function evaluation point and the objective function true value into a data pair;

[0017] The data pairs are added to the existing data set to expand the training data set.

[0018] In one embodiment, the step of improving the initial Gaussian process proxy model based on the training data set and the time decay factor to obtain a target optimization algorithm includes:

[0019] When sampling for the first time, multiple acquisition functions are evaluated according to the updated Gaussian process proxy model to obtain gain values ​​of multiple acquisition functions;

[0020] For each subsequent sampling, the updated Gaussian process proxy model is used to traverse the historical candidate points and update the historical gains of each acquisition function.

[0021] Until the number of iterations reaches the preset number of iterations or meets the convergence condition, the target optimization algorithm is obtained.

[0022] In one embodiment, the step of obtaining the objective function evaluation point by combining the initial Gaussian process proxy model with the acquisition function and the hyperparameters corresponding to the acquisition function includes:

[0023] Selecting an acquisition function according to the probability and updating a hyperparameter corresponding to the acquisition function, wherein the acquisition function includes an expected improvement, an improvement probability, and an upper confidence bound;

[0024] Generate multiple experimental points through LHS;

[0025] Based on the experimental points, the initial Gaussian process surrogate model combines the acquisition function and the hyperparameters corresponding to the acquisition function to obtain the evaluation points of the objective function.

[0026] In one embodiment, the step of obtaining the evaluation points of the objective function by combining the initial Gaussian process surrogate model with the acquisition function and the hyperparameters corresponding to the acquisition function based on the experimental points includes:

[0027] Calculate based on the experimental points to obtain a corresponding plurality of acquisition function values;

[0028] Based on the historical gain of the acquisition function corresponding to the plurality of acquisition function values, obtain the probability corresponding to the experimental points;

[0029] Select experimental points for evaluation based on the probability to obtain the evaluation points of the objective function.

[0030] In one embodiment, the step of generating a candidate point set by LHS includes:

[0031] Obtain the dimension of the parameter space;

[0032] Determine the number of candidate points according to the dimension of the parameter space;

[0033] According to the dimension of the parameter space, obtain the partition interval;

[0034] According to the partition interval and according to a preset rule, obtain a candidate point set.

[0035] In addition, to achieve the above object, the present application also proposes a Bayesian algorithm optimization device based on historical information, and the Bayesian algorithm optimization device based on historical information includes:

[0036] An initial module, configured to obtain an initial point set;

[0037] A construction module, configured to construct an initial Gaussian process surrogate model according to the initial point set;

[0038] An acquisition module, configured to obtain an acquisition function, the hyperparameters corresponding to the acquisition function, and a time decay factor based on the candidate point set generated by LHS;

[0039] A processing module, configured to obtain the evaluation points of the objective function in the candidate point set by combining the initial Gaussian process surrogate model with the acquisition function and the hyperparameters corresponding to the acquisition function;

[0040] A result module, configured to improve the initial Gaussian process surrogate model based on the evaluation points of the objective function and the time decay factor to obtain a target optimization algorithm.

[0041] In addition, to achieve the above-mentioned purpose, the present application also proposes a medium, which is a computer-readable medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the Bayesian algorithm optimization method based on historical information as described above are implemented.

[0042] In addition, to achieve the above-mentioned purpose, the present application also provides a computer program product, which includes a computer program, and when the computer program is executed by a processor, the steps of the Bayesian algorithm optimization method based on historical information as described above are implemented.

[0043] This application obtains an initial point set, constructs an initial Gaussian process proxy model based on the initial point set, generates a candidate point set through LHS, obtains the acquisition function, the hyperparameters corresponding to the acquisition function, and the time decay factor, obtains the target function evaluation point in the candidate point set through the initial Gaussian process proxy model combined with the acquisition function and the hyperparameters corresponding to the acquisition function, improves the initial Gaussian process proxy model based on the target function evaluation point and the time decay factor, and obtains the target optimization algorithm. By continuously updating the GP model, combining the dynamic acquisition function and its hyperparameters with the time decay factor, iteratively optimizes the target function evaluation point, and improves the optimization efficiency and accuracy of the Bayesian algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0045] Figure 1 This is a flow chart of the first embodiment of the Bayesian algorithm optimization method based on historical information of this application;

[0046] Figure 2 This is a comparison diagram of the effects of the first embodiment of the Bayesian algorithm optimization method based on historical information of this application;

[0047] Figure 3 This is a flow chart of the second embodiment of the Bayesian algorithm optimization method based on historical information of the present application;

[0048] Figure 4 This is a flowchart of the third embodiment of the Bayesian algorithm optimization method based on historical information of the present application;

[0049] Figure 5 This is a schematic diagram of the module structure of the Bayesian algorithm optimization device based on historical information in an embodiment of the present application;

[0050] Figure 6Schematic diagram of the device structure of the hardware operating environment involved in the Bayesian algorithm optimization method based on historical information in the embodiment of the present application.

[0051] The purpose, features and advantages of this application will be further described in conjunction with the embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0052] It should be understood that the specific embodiments described herein are only used to explain the technical solutions of the present application and are not used to limit the present application.

[0053] In order to better understand the technical solution of the present application, a detailed description will be given below in conjunction with the accompanying drawings and specific implementation methods.

[0054] The main solution of the embodiment of the present application is: by obtaining an initial point set, constructing an initial Gaussian process proxy model based on the initial point set, generating a candidate point set through LHS, obtaining the acquisition function, the hyperparameters corresponding to the acquisition function, and the time decay factor, combining the acquisition function and the hyperparameters corresponding to the acquisition function through the initial Gaussian process proxy model, obtaining the objective function evaluation point in the candidate point set, improving the initial Gaussian process proxy model based on the objective function evaluation point and the time decay factor, and obtaining the target optimization algorithm.

[0055] Based on this, the embodiment of the present application provides a Bayesian algorithm optimization method based on historical information, referring to Figure 1 , Figure 1 This is a flow chart of the first embodiment of the Bayesian algorithm optimization method based on historical information of the present application.

[0056] In this embodiment, the Bayesian algorithm optimization method based on historical information includes steps S10 to S50, which is specifically applied to the three-dimensional parameter optimization of plasma surface high-harmonic generation (SHHG) ellipticity, and Angelθ, Inclination Angleα and Preplasma Length L are selected. pre These three variables are scaled to the 0.1 interval within their respective value ranges for experiments, and the ellipticity of the attosecond pulse is used as the objective function for optimization:

[0057] Step S10, obtaining an initial point set.

[0058] It should be noted that the selection of the initial point set is a crucial step. It not only affects the quality of the proxy model, but also determines the direction and efficiency of subsequent iterative optimization. Therefore, selecting an appropriate number of initial points that are evenly distributed in the parameter space is crucial to the success of the entire optimization process.

[0059] Furthermore, the number of initial point sets should be determined according to the characteristics of the specific problem. If there are too few initial points, the constructed Gaussian process (GP) proxy model will be too simple and unable to accurately capture the true form of the objective function, thus affecting the effect of subsequent optimization. This is especially true in high-dimensional parameter spaces, because a small number of samples can hardly fully reflect the complex spatial structure, which can easily lead to local optimal solutions or unstable optimization paths. On the contrary, if there are too many initial points, the computational cost will increase, especially when it is very expensive to evaluate the objective function each time, such as when it involves complex physical simulations or long-running experimental designs. In addition, too many initial points will cover too many known areas, which is not conducive to finding the global optimal solution.

[0060] Furthermore, in order to ensure that the initial point set can effectively represent the entire parameter space, it should be distributed as evenly as possible. Even distribution means that the points in each dimension should be relatively evenly spread out to avoid some areas being too dense while other areas are almost not sampled. In this embodiment, the number of initial point sets is 1 / 5 of the total number of experiments. This number of sampled initial point data helps the proxy model learn the data features of different locations more comprehensively, thereby providing a more reliable basis for subsequent optimization.

[0061] Step S20, constructing an initial Gaussian process proxy model according to the initial point set.

[0062] It should be noted that GP is a nonparametric regression model that assumes that any finite number of random variables follow a multivariate normal distribution. The core of GP lies in its covariance function (also called kernel function), which defines a similarity measure between different input points. Commonly used kernel functions include square exponential kernel, Matérn kernel, etc. Selecting a suitable kernel function and adjusting its hyperparameters (such as length scale, signal variance, and noise variance) are crucial to capturing the characteristics of the objective function. When constructing a GP model, the above initial point set is used as training data to fit an initial GP model. This process involves solving the maximum likelihood estimate or using other methods to determine the optimal kernel function hyperparameters. After training, the GP model can predict the objective function value and its uncertainty for new input points, providing a basis for the selection of the acquisition function.

[0063] Step S30, generating a candidate point set through LHS, obtaining an acquisition function, a hyperparameter corresponding to the acquisition function, and a time attenuation factor.

[0064] It should be noted that in order to overcome the dimensionality curse problem caused by traditional grid search or random sampling methods in high-dimensional parameter space, Latin Hypercube Sampling (LHS) technology is used to generate a set of candidate points. LHS is a statistical method that ensures the uniform distribution of the selected sample points in the entire parameter space while reducing the number of samples required, thereby improving the efficiency and representativeness of sampling.

[0065] Further, step S30 includes: obtaining the dimension of the parameter space, and determining the number of candidate points according to the dimension of the parameter space. Specifically, first, it is necessary to clarify the dimension of the parameter space, that is, the number of independent variables involved in the optimization. High-dimensional parameter space means a larger search space and higher complexity, which puts higher requirements on the sampling strategy. For a given parameter space dimension, it is crucial to select an appropriate number of candidate points. Generally, the number of candidate points should be large enough to ensure good coverage of the parameter space, but not too large to waste computing resources. In practice, the specific selection often depends on the specific optimization problem, the available computing resources, and the requirements for exploration efficiency. The candidate points can not only provide preliminary information about the structure of the parameter space, but also be used to construct an initial Gaussian process (GP) proxy model. Then, according to the parameter space dimension, a partition interval is obtained, and according to the partition interval according to the preset rules, a candidate point set is obtained. Specifically, for each parameter space dimension, its value range is evenly divided into several equal probability intervals. Assuming that n candidate points need to be generated, the value range of each parameter will be divided into n equal parts. This ensures that each parameter is representative at different levels, thereby improving the coverage and diversity of the sample. In each dimension, a value is randomly selected from a pre-divided equal probability interval as a sample point in that dimension. This process is applied independently to each parameter dimension, and finally a set of candidate points is obtained.

[0066] It should be noted that the above elements jointly determine the efficiency of algorithm optimization. A variety of acquisition functions are used to guide the search process, including expected improvement (EI), probability of improvement (PI) and upper confidence bound (UCB). Each acquisition function has its own characteristics: EI measures the expected value of improvement that a new point may bring compared to the current optimal solution, PI calculates the probability that a new point exceeds the current optimal value, and UCB combines the predicted mean and uncertainty to encourage exploration of uncertain areas. The performance of the acquisition function depends on a set of hyperparameters, such as length scale, signal variance, and noise variance. These hyperparameters are not fixed but dynamically updated in each iteration. By combining the historical gain perception mechanism, the hyperparameters are adaptively adjusted according to the past evaluation results to better meet the needs of different stages. The time decay factor is introduced to control the importance of historical information and gradually reduce the weight of the long-term evaluation results as the number of iterations increases.

[0067] Step S40, obtaining the target function evaluation point in the candidate point set by combining the initial Gaussian process proxy model with the acquisition function and the hyperparameters corresponding to the acquisition function.

[0068] It should be noted that the hyperparameters of each acquisition function are dynamically adjusted, and the trend of the entire iterative process is comprehensively considered to gradually reduce the hyperparameter values ​​with the iteration. Then, the probabilities of multiple suggested experimental points are calculated based on the updated acquisition function, and the next objective function evaluation point is selected based on the probability. This selected point will be evaluated by the actual objective function, and its result will be added to the training data set to update the GP model.

[0069] Step S50, improving the initial Gaussian process proxy model based on the objective function evaluation point and the time decay factor to obtain a target optimization algorithm.

[0070] It should be noted that as the iteration progresses, each new evaluation point selected will be evaluated by the actual objective function and its results will be added to the training data set. These new evaluation points not only provide richer information about the objective function, but also help the GP model gradually approach the true objective function form. In order to ensure that the model can adapt to the changing data distribution, the GP model needs to be retrained after each iteration to update its parameters and improve the prediction accuracy. In this process, the time decay factor is introduced to adjust the importance of historical information. As time goes by, the value of the time decay factor gradually decreases, so that the most recent evaluation results are given higher importance, while the influence of earlier results gradually weakens. The GP proxy model is improved by the above method, and the target optimization algorithm is obtained. This algorithm continuously updates the GP model according to the latest evaluation results and the time decay factor in each iteration, so as to better capture the true characteristics of the objective function.

[0071] like Figure 2 From the comparison of the results shown in the figure, we can see that HAD-BO tends to converge during the iteration process and shows a more stable distribution. The ellipticity approaches the maximum value more quickly and stably. However, the original Bayesian optimization lacks clear regularity and optimization trend.

[0072] This embodiment provides a Bayesian algorithm optimization method based on historical information, by obtaining an initial point set, constructing an initial Gaussian process proxy model according to the initial point set, generating a candidate point set through LHS, obtaining an acquisition function, a hyperparameter corresponding to the acquisition function, and a time decay factor, combining the acquisition function and the hyperparameter corresponding to the acquisition function through the initial Gaussian process proxy model, obtaining an objective function evaluation point in the candidate point set, improving the initial Gaussian process proxy model based on the objective function evaluation point and the time decay factor, and obtaining an objective optimization algorithm. By continuously updating the GP model, combining the dynamic acquisition function and its hyperparameters with the time decay factor, iteratively optimizing the objective function evaluation point, the Bayesian algorithm optimization efficiency and accuracy are improved.

[0073] Based on the first embodiment of the present application, in the second embodiment of the present application, the same or similar contents as those in the above-mentioned embodiment 1 can refer to the above introduction, and will not be repeated later. Figure 3 The Bayesian algorithm optimization method based on historical information step S40 also includes steps S401 to S403:

[0074] Step S401, selecting and updating an acquisition function and hyperparameters corresponding to the acquisition function according to probability, wherein the acquisition function includes expected improvement, improvement probability and confidence upper bound.

[0075] It should be noted that at the beginning of each iteration, the hyperparameters of the acquisition function are updated according to the probability, and these updated parameters are used to guide the selection of new evaluation points. The above probabilities are selected by setting rules to specify when to use which acquisition function is more likely to be suitable for the current scenario. For example, EI measures the expected value of the improvement that a new point may bring compared to the current optimal solution, PI calculates the probability that a new point exceeds the current optimal value, and UCB combines the predicted mean and uncertainty. In this way, flexibility and adaptability can be maintained in a constantly changing optimization environment, ultimately achieving faster convergence speed and higher optimization accuracy.

[0076] Step S402, generating multiple experimental points through LHS.

[0077] It should be noted that the updated acquisition function is used to generate multiple candidate experimental points. Specifically, for each acquisition function, multiple recommended experimental points are proposed in the parameter space, and weighted selection is performed according to the probability provided by each acquisition function.

[0078] Step S403, obtaining the objective function evaluation point according to the experimental point through the initial Gaussian process proxy model combined with the acquisition function and the hyper parameters corresponding to the acquisition function.

[0079] It should be noted that the hyperparameters of each acquisition function are dynamically optimized, and the trend of the entire iterative process is comprehensively considered so that the hyperparameter values ​​are gradually reduced with the iteration. Then, the probabilities of multiple experimental points are calculated based on the updated acquisition function, and the next objective function evaluation point is selected based on the probability. This selected point will be evaluated by the actual objective function, and its result will be added to the training data set. In the specific operation, the optimal point of the acquisition function is obtained, and the formula is expressed as:

[0080]

[0081] Among them, u i (x) is the value of the i-th acquisition function, X Predict is the candidate point set. Then the historical gains of multiple acquisition functions are combined through the probability formula:

[0082]

[0083] in is the gain value of each acquisition function, is the historical gain update function. The target function is used to actually sample the selected points to obtain the sampled values, and then the sampled values ​​are added to the training data set.

[0084] Further, step S403 also includes: performing calculations according to the experimental points to obtain the corresponding multiple acquisition function values, and then obtaining the probability corresponding to the experimental points according to the historical gains of the acquisition functions corresponding to the multiple acquisition function values, and finally selecting the experimental points for evaluation according to the above probabilities to obtain the objective function evaluation points. Specifically, the experimental points are calculated according to the corresponding standards by the acquisition function to obtain the corresponding acquisition function values. Next, the probability of each experimental point is calculated according to the historical gains of the acquisition functions corresponding to each acquisition function value, which usually involves weighted averaging of different acquisition function values ​​or other comprehensive evaluation methods to obtain an overall score. The higher the probability, the more worthy the experimental point is to be selected as the next evaluation point. This mechanism not only improves the sampling efficiency, but also increases the possibility of finding the global optimal solution. Finally, the experimental points are selected according to the probability for actual objective function evaluation. The probability depends on the historical gains of each acquisition function. A time decay factor is introduced in the selection process to control the importance of historical information. As time goes by, the influence of distant historical evaluation results gradually decreases, so that the most recent evaluation results are given higher importance, while the influence of earlier results gradually weakens. This mechanism helps the model to more effectively utilize the latest information while avoiding being misled by early inaccurate estimates. Once a new evaluation point is selected and the actual objective function is evaluated, the evaluation results are added to the training data set to retrain the GP model. This update process ensures that the GP model always reflects the latest data, thereby improving the accuracy of subsequent predictions.

[0085] This embodiment generates multiple experimental points by dynamically selecting and updating the acquisition function and its hyperparameters, and uses the initial GP model for evaluation to accelerate convergence and improve accuracy.

[0086] Based on the first embodiment of the present application, in the third embodiment of the present application, the same or similar contents as those in the first embodiment can be referred to the above description, and will not be described in detail later. Figure 4 The Bayesian algorithm optimization method based on historical information step S50 also includes steps S501 to S503:

[0087] Step S501, obtaining the true value of the objective function according to the objective function evaluation point.

[0088] It should be noted that once a new evaluation point (i.e., objective function evaluation point) is selected, the next step is to perform the actual objective function evaluation on it. This step is especially important for expensive objective functions, because each evaluation may require a lot of resources, such as time, computing power, or experimental costs. After the evaluation is completed, the true value of the objective function corresponding to that point will be obtained. This value directly reflects the performance of the objective function under the parameter setting, and in the physical experiment design, it represents the measurement result of a certain physical quantity. By obtaining these true values ​​of the objective function, we can more accurately understand the true form and characteristics of the objective function.

[0089] Step S502: Expand the training data set based on the objective function evaluation points and the objective function true value.

[0090] It should be noted that after obtaining the objective function evaluation points and their corresponding objective function values, these data can be combined into new training data and added to the training data set. The training data set consists of input (i.e., the parameter settings of the candidate points) and output (i.e., the corresponding objective function values) and is used to train and update the GP agent model.

[0091] Further, step S502 also includes: according to the objective function evaluation point, using the objective function to calculate, and obtaining the objective function true value corresponding to the objective function evaluation point; combining the objective function evaluation point and the objective function true value into a data pair; adding the data pair to the existing data set to expand the training data set. Specifically, for each candidate point, use the acquisition function (such as expected improvement EI, probability improvement PI and confidence upper bound UCB) to guide the selection of the next evaluation point. The acquisition function measures the value of potential evaluation points and helps determine which point is most likely to bring significant improvement. For example, EI tends to select points that may bring the greatest improvement; PI prefers points with a higher probability of exceeding the current best value; and UCB is more focused on exploring uncertain but potential areas. Once a new evaluation point is selected, the actual objective function evaluation is performed on it to obtain the corresponding objective function true value. As the iteration proceeds, multiple such training data pairs will be continuously generated. These data pairs are summarized to construct an updated training data set. Specifically, all existing training data pairs are collected to form a data set containing all evaluated candidate points and their corresponding objective function true values. This training data set is used to train and update the GP agent model. The GP model assumes that any finite number of random variables follow a multivariate normal distribution and can predict the objective function value and its uncertainty for new input points. With the addition of more evaluation points, the GP model gradually approaches the true objective function, improves the accuracy of the prediction, and provides a more reliable basis for the selection of subsequent acquisition functions.

[0092] Step S503, improving the initial Gaussian process proxy model based on the training data set and the time decay factor to obtain a target optimization algorithm.

[0093] It should be noted that the time decay factor is introduced to adjust the importance of historical information. As time goes by, the most recent evaluation results will receive higher importance, while the impact of earlier results will gradually weaken.

[0094] Further, step S503 includes: when sampling for the first time, multiple acquisition functions are evaluated according to the initial Gaussian process proxy model to obtain gain values ​​of multiple acquisition functions. Specifically, when sampling for the first time, multiple acquisition functions are evaluated simultaneously. For each acquisition function, its mean value at each candidate point on the Gaussian process proxy model is calculated, and the gain value is determined based on these values. The gain value reflects the importance of each acquisition function in the current optimization stage and provides a basis for subsequent selection. At the same time, it is allowed to consider the results of multiple acquisition functions simultaneously in the same iteration. This means that, when sampling for the first time, multiple recommended experimental points from different acquisition functions may be comprehensively considered. Doing so can make full use of the advantages of different acquisition functions and achieve a more comprehensive and balanced search strategy. In addition, the time decay factor is introduced to control the importance of historical information. Although the impact is small at the first sampling, it lays the foundation for dynamic adjustment in subsequent iterations.

[0095] Furthermore, step S503 also includes: for each subsequent sampling, the historical candidate points are traversed through the updated Gaussian process proxy model, and the historical gains of each acquisition function are updated; until the number of iterations reaches the preset number of iterations or meets the convergence condition, the target optimization algorithm is obtained. Specifically, from the end of the first sampling to the preset number of iterations or the convergence condition is met, at the beginning of each iteration, all previously evaluated points, i.e., historical candidate points, are reviewed. Through these points, the algorithm can obtain the best historical points, which provides a reference for selecting the next potential experimental point. Then the current version of the GP proxy model is used to predict the historical candidate points to obtain their objective function mean. This predicted value reflects the estimation of the performance of each point based on the existing information, helping to determine which areas may hide better solutions. The time decay factor k is introduced to gradually reduce the weight of the long-term evaluation results. The purpose of this is to make the algorithm pay more attention to recent information while not completely ignoring past data. For each historical point x j , calculate its new gain value, and the update function is expressed as:

[0096]

[0097] where g i is the base value of historical gain, k is the time attenuation factor, μ(x j ) is the target value mean, Represents the gain value of the jth historical point after the i-th iteration. As the number of iterations increases, the value of k gradually decreases, so that the influence of early evaluations gradually weakens. According to the latest historical gain and GP model output, a dynamic acquisition strategy is used to determine the location of the next experimental point. In this process, the algorithm selects several different acquisition functions (such as expected improvement EI, probability improvement PI, and confidence upper bound UCB), and dynamically updates its hyperparameters for each acquisition function in each iteration. The selected new experimental point will be actually evaluated to obtain its true objective function value, and this pair of input-output data will be added to the training set. Subsequently, the GP proxy model is retrained with the updated data set to better fit the actual situation. After each iteration, check whether the preset maximum number of iterations has been reached or whether a certain convergence criterion has been met (for example, the change in the objective function value in several consecutive iterations is very small). If any of the conditions is met, the iteration is stopped and the GP proxy model at this time is output, that is, the target optimization algorithm.

[0098] This embodiment obtains the objective function value based on the objective function evaluation point, forms a training data set in combination with the candidate point set, and uses the time decay factor to update the initial GP model to achieve efficient optimization and improve the convergence speed and accuracy.

[0099] This application also provides a Bayesian algorithm optimization device based on historical information, please refer to Figure 5 , the device comprises:

[0100] The initial module 10 is used to obtain an initial point set.

[0101] The construction module 20 is used to construct an initial Gaussian process proxy model according to the initial point set.

[0102] The acquisition module 30 is used to generate a candidate point set through LHS, and obtain an acquisition function, a hyperparameter corresponding to the acquisition function, and a time attenuation factor.

[0103] The processing module 40 is used to obtain the target function evaluation point in the candidate point set by combining the acquisition function and the hyperparameters corresponding to the acquisition function through the initial Gaussian process proxy model.

[0104] The result module 50 is used to improve the initial Gaussian process proxy model based on the objective function evaluation point and the time decay factor to obtain the target optimization algorithm.

[0105] The Bayesian algorithm optimization device based on historical information provided by the present application adopts the Bayesian algorithm optimization method based on historical information in the above-mentioned embodiment, which can solve the technical problem of how to improve the efficiency of Bayesian algorithm optimization. Compared with the prior art, the beneficial effects of the Bayesian algorithm optimization device based on historical information provided by the present application are the same as the beneficial effects of the Bayesian algorithm optimization method based on historical information provided by the above-mentioned embodiment, and the other technical features in the Bayesian algorithm optimization device based on historical information are the same as the features disclosed in the above-mentioned embodiment method, which will not be repeated here.

[0106] In one embodiment, the result module 50 is also used to obtain the true value of the objective function based on the objective function evaluation point; expand the training data set based on the objective function evaluation point and the true value of the objective function; improve the initial Gaussian process proxy model based on the training data set and the time attenuation factor to obtain the target optimization algorithm.

[0107] In one embodiment, the result module 50 is also used to calculate using the objective function according to the objective function evaluation point to obtain the objective function true value corresponding to the objective function evaluation point; combine the objective function evaluation point and the objective function true value into a data pair; add the data pair to the existing data set to expand the training data set.

[0108] In one embodiment, the result module 50 is also used to evaluate multiple acquisition functions according to the initial Gaussian process proxy model when sampling for the first time, and obtain gain values ​​of multiple acquisition functions; after the first sampling, obtain the historical best point by traversing the historical candidate points; predict the target value mean of the historical points through the initial Gaussian process proxy model, and update the historical gain; until the number of iterations reaches the preset number of iterations or meets the convergence conditions, the target optimization algorithm is obtained.

[0109] In one embodiment, the processing module 40 is also used to select and update the acquisition function and the hyperparameters corresponding to the acquisition function, where the acquisition function includes expected improvement, improvement probability and confidence upper bound; generate multiple experimental points according to the acquisition function and the hyperparameters corresponding to the acquisition function; and obtain the objective function evaluation point according to the experimental points through the initial Gaussian process proxy model.

[0110] In one embodiment, the processing module 40 is also used to perform calculations based on the experimental points to obtain corresponding acquisition function values; obtain the priorities corresponding to the experimental points based on the acquisition function values; and select experimental points with priorities that are preset values ​​for evaluation to obtain objective function evaluation points.

[0111] In one embodiment, the acquisition module 30 is further used to obtain the parameter space dimension; determine the number of candidate points according to the parameter space dimension; obtain the partition interval according to the parameter space dimension; and obtain the candidate point set based on the initial point set according to the partition interval according to the preset rules.

[0112] The present application provides a Bayesian algorithm optimization device based on historical information, and the Bayesian algorithm optimization device based on historical information includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the Bayesian algorithm optimization method based on historical information in the above-mentioned embodiment one.

[0113] Reference below Figure 6 , which shows a schematic diagram of the structure of a Bayesian algorithm optimization device based on historical information suitable for implementing the embodiment of the present application. The Bayesian algorithm optimization device based on historical information in the embodiment of the present application may include but is not limited to mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), etc., and fixed terminals such as digital TVs, desktop computers, etc. Figure 6 The Bayesian algorithm optimization device based on historical information shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present application.

[0114] like Figure 6As shown, the Bayesian algorithm optimization device based on historical information may include a processing device 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM: Read Only Memory) 1002 or a program loaded from a storage device 1003 to a random access memory (RAM: Random Access Memory) 1004. Various programs and data required for the operation of the Bayesian algorithm optimization device based on historical information are also stored in RAM1004. The processing device 1001, ROM1002, and RAM1004 are connected to each other via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to the I / O interface 1006: input devices 1007 including, for example, a touch screen, a touchpad, a keyboard, a mouse, an image sensor, a microphone, an accelerometer, a gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; storage devices 1003 including, for example, a magnetic tape, a hard disk, etc.; and communication devices 1009. The communication device 1009 can allow the Bayesian algorithm optimization device based on historical information to communicate wirelessly or wired with other devices to exchange data. Although the Bayesian algorithm optimization device based on historical information with various systems is shown in the figure, it should be understood that it is not required to implement or have all the systems shown. More or fewer systems may be implemented or have alternatively.

[0115] In particular, according to the embodiments disclosed in the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiments disclosed in the present application include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network through a communication device, or installed from a storage device 1003, or installed from a ROM 1002. When the computer program is executed by the processing device 1001, the above-mentioned functions defined in the method of the embodiment disclosed in the present application are executed.

[0116] The Bayesian algorithm optimization device based on historical information provided by the present application adopts the Bayesian algorithm optimization method based on historical information in the above-mentioned embodiment, which can solve the technical problem of how to improve the efficiency of Bayesian algorithm optimization. Compared with the prior art, the beneficial effects of the Bayesian algorithm optimization device based on historical information provided by the present application are the same as the beneficial effects of the Bayesian algorithm optimization method based on historical information provided by the above-mentioned embodiment, and the other technical features in the Bayesian algorithm optimization device based on historical information are the same as the features disclosed in the method of the previous embodiment, which will not be repeated here.

[0117] It should be understood that the various parts disclosed in this application can be implemented by hardware, software, firmware or a combination thereof. In the description of the above embodiments, specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0118] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0119] The present application provides a computer-readable medium having computer-readable program instructions (ie, computer programs) stored thereon, and the computer-readable program instructions are used to execute the Bayesian algorithm optimization method based on historical information in the above-mentioned embodiment.

[0120] The computer-readable medium provided in the present application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, systems or devices, or any combination of the above. More specific examples of computer-readable media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In this embodiment, the computer-readable medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, system or device. The program code contained on the computer-readable medium may be transmitted using any appropriate medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination of the above.

[0121] The computer-readable medium may be included in the Bayesian algorithm optimization device based on historical information; or may exist independently without being assembled into the Bayesian algorithm optimization device based on historical information.

[0122] The computer-readable medium carries one or more programs. When the one or more programs are executed by the Bayesian algorithm optimization device based on historical information, the Bayesian algorithm optimization device based on historical information can be written in one or more programming languages ​​or a combination thereof to write computer program codes for performing the operations of the present application. The programming languages ​​include object-oriented programming languages, such as Java, Smalltalk, C++, and conventional procedural programming languages, such as "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as an independent software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computer (for example, using an Internet service provider to connect through the Internet).

[0123] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present application. In this regard, each square box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, the program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two square boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0124] The modules involved in the embodiments described in this application may be implemented by software or hardware, wherein the name of the module does not constitute a limitation on the unit itself in some cases.

[0125] The readable medium provided by the present application is a computer-readable medium, which stores computer-readable program instructions (i.e., computer programs) for executing the above-mentioned Bayesian algorithm optimization method based on historical information, and can solve the technical problem of how to improve the efficiency of Bayesian algorithm optimization. Compared with the prior art, the beneficial effects of the computer-readable medium provided by the present application are the same as the beneficial effects of the Bayesian algorithm optimization method based on historical information provided by the above-mentioned embodiment, and will not be repeated here.

[0126] The present application also provides a computer program product, including a computer program, which implements the steps of the Bayesian algorithm optimization method based on historical information as described above when executed by a processor.

[0127] The computer program product provided by the present application can solve the technical problem of how to improve the efficiency of Bayesian algorithm optimization. Compared with the prior art, the beneficial effects of the computer program product provided by the present application are the same as the beneficial effects of the Bayesian algorithm optimization method based on historical information provided by the above embodiment, which will not be repeated here.

[0128] The above descriptions are only some embodiments of the present application, and are not intended to limit the patent scope of the present application. All equivalent structural changes made using the contents of the present application specification and drawings under the technical concept of the present application, or direct / indirect applications in other related technical fields are included in the patent protection scope of the present application.

Claims

1. A Bayesian algorithm optimization method based on historical information, characterized in that: The method comprises: Get the initial point set; Constructing an initial Gaussian process proxy model according to the initial point set; Generate a candidate point set through LHS, obtain an acquisition function, a hyperparameter corresponding to the acquisition function, and a time decay factor; Obtaining an objective function evaluation point in a candidate point set by combining the initial Gaussian process proxy model with the acquisition function and a hyperparameter corresponding to the acquisition function; The initial Gaussian process proxy model is improved based on the objective function evaluation point and the time decay factor to obtain a target optimization algorithm.

2. The method according to claim 1, characterized in that The step of improving the initial Gaussian process proxy model based on the objective function evaluation point and the time decay factor to obtain a target optimization algorithm includes: According to the objective function evaluation point, a true value of the objective function is obtained; Expanding the training data set based on the objective function evaluation point and the objective function true value; The initial Gaussian process proxy model is improved based on the training data set and the time decay factor to obtain a target optimization algorithm.

3. The method according to claim 2, characterized in that The step of expanding the training data set based on the objective function evaluation point and the objective function true value comprises: According to the objective function evaluation point, the objective function is used to perform calculation to obtain the true value of the objective function corresponding to the objective function evaluation point; Combining the objective function evaluation point and the objective function true value into a data pair; The data pairs are added to the existing data set to expand the training data set.

4. The method according to claim 2, characterized in that The step of improving the initial Gaussian process proxy model based on the training data set and the time decay factor to obtain a target optimization algorithm includes: When sampling for the first time, multiple acquisition functions are evaluated according to the updated Gaussian process proxy model to obtain gain values ​​of the multiple acquisition functions; for each subsequent sampling, the historical candidate points are traversed through the updated Gaussian process proxy model to update the historical gain of each acquisition function; Until the number of iterations reaches the preset number of iterations or meets the convergence condition, the target optimization algorithm is obtained.

5. The method according to claim 1, characterized in that The step of obtaining the objective function evaluation point in the candidate point set by combining the initial Gaussian process proxy model with the acquisition function and the hyperparameters corresponding to the acquisition function comprises: Selecting an acquisition function according to the probability and updating a hyperparameter corresponding to the acquisition function, wherein the acquisition function includes an expected improvement, an improvement probability, and an upper confidence bound; Generate multiple experimental points through LHS; According to the experimental point, the objective function evaluation point is obtained by combining the acquisition function and the hyperparameters corresponding to the acquisition function through the initial Gaussian process proxy model.

6. The method according to claim 5, characterized in that The step of obtaining the objective function evaluation point according to the experimental point by combining the initial Gaussian process proxy model with the acquisition function and the hyperparameters corresponding to the acquisition function comprises: Calculating according to the experimental points to obtain corresponding multiple acquisition function values; Obtaining the probability corresponding to the experimental point according to the historical gains of the acquisition function corresponding to the plurality of acquisition function values; An experimental point is selected for evaluation according to the probability to obtain an evaluation point of the objective function.

7. The method according to claim 1, characterized in that The step of generating a candidate point set through LHS includes: Get the parameter space dimension; Determining the number of candidate points according to the dimension of the parameter space; According to the parameter space dimension, a partition interval is obtained; According to the division intervals and preset rules, a candidate point set is obtained.

8. A Bayesian algorithm optimization device based on historical information, characterized in that: The device comprises: Initial module, used to obtain the initial point set; A construction module, used to construct an initial Gaussian process proxy model according to the initial point set; An acquisition module, used to generate a candidate point set through LHS, and acquire an acquisition function, a hyperparameter corresponding to the acquisition function, and a time decay factor; A processing module, configured to obtain an objective function evaluation point in a candidate point set by combining the initial Gaussian process proxy model with the acquisition function and a hyperparameter corresponding to the acquisition function; A result module is used to improve the initial Gaussian process proxy model based on the objective function evaluation point and the time decay factor to obtain a target optimization algorithm.

9. A Bayesian algorithm optimization device based on historical information, characterized in that: The device includes: a memory, a processor, and a Bayesian algorithm optimization program based on historical information stored in the memory and running on the processor, wherein the Bayesian algorithm optimization program based on historical information is configured to implement the steps of the Bayesian algorithm optimization method based on historical information as described in any one of claims 1-7.

10. A storage medium, characterized in that: The storage medium stores a Bayesian algorithm optimization program based on historical information, and when the Bayesian algorithm optimization program based on historical information is executed by the processor, the steps of the Bayesian algorithm optimization method based on historical information as described in any one of claims 1 to 7 are implemented.

Images